article-workflow-final-review

Article Workflow Final Review

这是流程类 Skill，默认仅在用户显式调用 $article-workflow-final-review 时运行。

This is Phase 6 of the article-workflow-* skill series, corresponding to "Final Refinement and Reader Testing" in the article optimization workflow overview.

Goal: Final refinement before publication — including title, opening, ending, credibility, reader experience, and illustration placement. AI performs final editorial review and inserts confirmed illustrations; the author makes the final publish decision. Do not rewrite the entire narrative arc.

Trigger Scenarios

Use this skill when the user asks to:

• Execute Phase 6 of the article optimization workflow
• Perform final review / pre-publish check / final review
• Generate .article-workflow/06-final-review/06-final-review.md
• Generate .article-workflow/06-final-review/final.md
• Insert confirmed screenshots, diagrams, or AI images into the final draft per the visual plan

Input and Output

Default inputs:

• .article-workflow/brief.md
• .article-workflow/style-bible.md
• .article-workflow/05-polish-rounds/polished-draft.md
• .article-workflow/04.6-visual-plan/visual-plan.md (if present)
• .article-workflow/04.6-visual-plan/asset-manifest.md (if present)
• .article-workflow/04.6-visual-plan/ai-image-briefs.md (if present and confirmed by author)
• .article-workflow/sources/proofs/ (if present)

Default outputs go into the Phase 6 directory:

• .article-workflow/06-final-review/06-final-review.md
• .article-workflow/06-final-review/final.md: Publishable version generated after the author confirms final edits, including confirmed illustration placements

If the user provides an article directory, first look for .article-workflow/ under that directory.

Core Principles

• Final refinement, not rewriting the narrative arc. Only address the most worthwhile pre-publish issues, confirmed illustration placements, title/opening/ending/credibility checks — do not reorganize the entire structure.
• Review first, then finalize. By default, write 06-final-review.md first; only write/update final.md when the author confirms final edits, or the user explicitly requests generating the final draft in the current round.
• Only insert illustrations the author has confirmed or the visual plan explicitly adopts. If the image file already exists, insert it directly; if the file hasn't been generated yet but visual-plan.md, asset-manifest.md, imgs/outline.md, or prompt files have explicitly defined the position and target filename, you can insert the planned path, but must list it as a pre-publish file confirmation item in the review report.
• Apply discretion to explanatory infographics. Don't insert every explanatory infographic just because a prompt/outline exists. Prioritize evidence screenshots and one core mechanism diagram that truly lowers the comprehension barrier; images that repeat the body text, weaken narrative pacing, or turn the article into slides should not be inserted — explain the reason.
• Evidence images and AI images must be clearly distinguished. Evidence screenshots retain their source attribution; AI images must be labeled as explanatory illustrations or atmosphere images, never disguised as evidence.
• Images cannot replace body-text argumentation. Screenshots and diagrams can only aid understanding or support facts; the body text must first make clear why the reader should look at the image.
• Better to omit than to clutter. If an image would interrupt reading, repeat the body text, have unclear evidential value, or lack a clear caption/alt text, don't insert it — and explain why in 06-final-review/06-final-review.md.
• Final quality comes from the author reading through. AI can do final editorial review, but cannot replace the author reading the piece end to end. After generating final.md, you must remind the author to read the entire piece; places that read awkwardly, don't sound like the author, or feel skippable should be addressed by extracting specific paragraphs and having a focused conversation with AI for localized edits — not by reopening a large-scope process.
• When encountering ambiguities that would affect the publish decision (title direction, whether to make a screenshot public, whether AI images are allowed, whether final is publishable), proactively ask the author; if prior materials are already clear, don't force questions.
• After writing outputs, you must launch a sub-agent review; the sub-agent only provides review comments, and the main agent decides whether to adopt them.